{"id":"f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7","title":"How Much Data SAE Training Needs","argument":"Authors Nikita Koriagin, Nikita Balagansky and Daniil Gavrilov study how much activation data sparse autoencoder (SAE) training needs when its total training budget is fixed. They report that quality improves as the activation buffer grows, then reaches a plateau: replaying that buffer is nearly as effective as using fresh activations throughout training. Beyond the plateau, they say, adding unique activations brings little benefit to reconstruction or interpretability. The authors describe this diversity-versus-repetition tradeoff with a scaling law for limited data and report testing it across model layers and sizes, SAE types, token budgets and activation sources. Their practical rule is to use the smallest buffer that reaches the quality plateau, then replay it for the remaining budget. In their experiments, this reduced activation storage by 8–64× with negligible quality loss. The result could help reduce a storage bottleneck in SAE training, but the reported savings and quality outcome are experimental findings, not a guarantee for other settings.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied abstract gives no numeric buffer sizes or per-condition results, so it does not show how the reported plateau or storage reduction varies across the tested settings.","speaker":{"name":"Koriagin, Nikita, Balagansky, Nikita, Gavrilov, Daniil","roleAtSource":"Article author"},"source":{"title":"Scaling Laws for SAE Training Data","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.07756","publishedAt":"2026-11-05","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors report that SAE quality improves as the activation buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","excerpt":"Quality improves rapidly as the buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"The authors say additional unique activations beyond the plateau bring little benefit to reconstruction or interpretability.","excerpt":"Beyond this point, additional unique activations bring little benefit to reconstruction or interpretability","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"The authors recommend choosing the smallest buffer that reaches the quality plateau and replaying it for the remaining training budget.","excerpt":"choose the smallest buffer that reaches the quality plateau and replay it for the remaining budget.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"In their experiments, the authors report reducing activation storage by 8–64× with negligible quality loss.","excerpt":"In our experiments, this reduces activation storage by $8\\text{--}64\\times$ with negligible quality loss.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"The authors report testing the tradeoff across dictionary sizes, token budgets, model layers and sizes, SAE types, and activation sources.","excerpt":"We validate it across dictionary sizes and token budgets, early, middle, and late layers of Llama-3.1, Qwen3 model sizes, and BatchTopK, TopK, and JumpReLU SAEs.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.07756","checkedAt":"2026-11-08T18:50:08.498Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does the quality plateau change when the activation source changes?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"How Much Data SAE Training Needs","summary":"Authors Nikita Koriagin, Nikita Balagansky and Daniil Gavrilov study how much activation data sparse autoencoder (SAE) training needs when its total training budget is fixed. They report that quality improves as the activation buffer grows, then reaches a plateau: replaying that buffer is nearly as effective as using fresh activations throughout training. Beyond the plateau, they say, adding unique activations brings little benefit to reconstruction or interpretability. The authors describe this diversity-versus-repetition tradeoff with a scaling law for limited data and report testing it across model layers and sizes, SAE types, token budgets and activation sources. Their practical rule is to use the smallest buffer that reaches the quality plateau, then replay it for the remaining budget. In their experiments, this reduced activation storage by 8–64× with negligible quality loss. The result could help reduce a storage bottleneck in SAE training, but the reported savings and quality outcome are experimental findings, not a guarantee for other settings.","question":"How does the quality plateau change when the activation source changes?","claims":[{"text":"The authors report that SAE quality improves as the activation buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","excerpt":"Quality improves rapidly as the buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","support":"source-stated"},{"text":"The authors say additional unique activations beyond the plateau bring little benefit to reconstruction or interpretability.","excerpt":"Beyond this point, additional unique activations bring little benefit to reconstruction or interpretability","support":"source-stated"},{"text":"The authors recommend choosing the smallest buffer that reaches the quality plateau and replaying it for the remaining training budget.","excerpt":"choose the smallest buffer that reaches the quality plateau and replay it for the remaining budget.","support":"source-stated"},{"text":"In their experiments, the authors report reducing activation storage by 8–64× with negligible quality loss.","excerpt":"In our experiments, this reduces activation storage by $8\\text{--}64\\times$ with negligible quality loss.","support":"source-stated"},{"text":"The authors report testing the tradeoff across dictionary sizes, token budgets, model layers and sizes, SAE types, and activation sources.","excerpt":"We validate it across dictionary sizes and token budgets, early, middle, and late layers of Llama-3.1, Qwen3 model sizes, and BatchTopK, TopK, and JumpReLU SAEs.","support":"source-stated"}],"limitations":["The supplied abstract gives no numeric buffer sizes or per-condition results, so it does not show how the reported plateau or storage reduction varies across the tested settings."],"mediaId":"1d462f6593cfc99d347bfb4bce73e1ae67fc01471754d1880422eedc33d5485e","correctionReason":null,"id":"f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7","publicationId":"1e4588f7-3a93-4627-8fec-168b41da258a","version":1,"sourceUrl":"https://arxiv.org/abs/2611.07756","sourceTitle":"Scaling Laws for SAE Training Data","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-05","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:00:59.872Z","createdAt":"2026-11-08T17:46:56.079Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"1d462f6593cfc99d347bfb4bce73e1ae67fc01471754d1880422eedc33d5485e","kind":"article","url":"https://arxiv.org/abs/2611.07756","sourceUrl":"https://arxiv.org/abs/2611.07756","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-08T17:46:15.101Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.07756","checkedAt":"2026-11-08T18:50:08.498Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"1d462f6593cfc99d347bfb4bce73e1ae67fc01471754d1880422eedc33d5485e","confidence":0.8700000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"4dbec366c1a85999662762ed6dfa7f499d2339596c9f68a3dd62eb6bf86ca590","decidedAt":"2026-11-08T17:46:15.672Z"},"provenance":{"textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observationId":"1d57713265820b7b1359184088c4e2708a09ba79c33cd54ea6e83f3edd361f33","retrievedAt":"2026-11-08T17:46:15.004Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"3a9c6344-5ffe-4bd1-871a-e4d888faca3e"},"reviewer":{"model":"gpt-6-luna","runId":"3ac55bec-dc3d-4715-822e-32476d5145d3"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T17:46:56.079Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Koriagin, Nikita","url":null,"kind":"unknown"},{"name":"Balagansky, Nikita","url":null,"kind":"unknown"},{"name":"Gavrilov, Daniil","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Koriagin, Nikita","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Koriagin, Nikita","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.07756","locator":"head/meta[name=author|citation_author]"},"assertionId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T18:50:08.498Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","sourceUrl":"https://arxiv.org/abs/2611.07756","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z"}},{"actor":{"name":"Balagansky, Nikita","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Balagansky, Nikita","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.07756","locator":"head/meta[name=author|citation_author]"},"assertionId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T18:50:08.498Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","sourceUrl":"https://arxiv.org/abs/2611.07756","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z"}},{"actor":{"name":"Gavrilov, Daniil","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gavrilov, Daniil","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.07756","locator":"head/meta[name=author|citation_author]"},"assertionId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T18:50:08.498Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","sourceUrl":"https://arxiv.org/abs/2611.07756","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z"}}],"participationCoverage":{"total":3,"returned":3,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.07756","snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7/source"}},"canonicalUrl":"https://escapemeridian.com/readings/how-much-data-sae-training-needs-f287900f","formats":{"html":"/readings/how-much-data-sae-training-needs-f287900f","markdown":"/readings/how-much-data-sae-training-needs-f287900f.md","json":"/readings/how-much-data-sae-training-needs-f287900f.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-interpretability","name":"Interpretability","explanation":"Research into how a model reaches its outputs, often by examining its internal workings.","referenceUrl":"https://developers.google.com/machine-learning/glossary#interpretability"},{"id":"term-scaling-laws","name":"Scaling laws","explanation":"Observed patterns in which model performance improves predictably as data, model size and computation grow.","referenceUrl":"https://en.wikipedia.org/wiki/Neural_scaling_law"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"},{"id":"buffer","name":"Buffer","kind":"organization","context":"Buffer makes social media management software and publishes company accounts of customer support, team practices and online publishing.","referenceUrls":["https://buffer.com/"],"path":"/entities/buffer","json":"/entities/buffer.json","markdown":"/entities/buffer.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7","revision":"1e4588f7-3a93-4627-8fec-168b41da258a","revisionKind":"publication-id","title":"How Much Data SAE Training Needs","url":"https://escapemeridian.com/readings/how-much-data-sae-training-needs-f287900f"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.07756","revision":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","revisionKind":"sha256-source-text","title":"Scaling Laws for SAE Training Data","url":"https://arxiv.org/abs/2611.07756"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7","title":"How Much Data SAE Training Needs","argument":"Authors Nikita Koriagin, Nikita Balagansky and Daniil Gavrilov study how much activation data sparse autoencoder (SAE) training needs when its total training budget is fixed. They report that quality improves as the activation buffer grows, then reaches a plateau: replaying that buffer is nearly as effective as using fresh activations throughout training. Beyond the plateau, they say, adding unique activations brings little benefit to reconstruction or interpretability. The authors describe this diversity-versus-repetition tradeoff with a scaling law for limited data and report testing it across model layers and sizes, SAE types, token budgets and activation sources. Their practical rule is to use the smallest buffer that reaches the quality plateau, then replay it for the remaining budget. In their experiments, this reduced activation storage by 8–64× with negligible quality loss. The result could help reduce a storage bottleneck in SAE training, but the reported savings and quality outcome are experimental findings, not a guarantee for other settings.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied abstract gives no numeric buffer sizes or per-condition results, so it does not show how the reported plateau or storage reduction varies across the tested settings.","speaker":{"name":"Koriagin, Nikita, Balagansky, Nikita, Gavrilov, Daniil","roleAtSource":"Article author"},"source":{"title":"Scaling Laws for SAE Training Data","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.07756","publishedAt":"2026-11-05","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors report that SAE quality improves as the activation buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","excerpt":"Quality improves rapidly as the buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"The authors say additional unique activations beyond the plateau bring little benefit to reconstruction or interpretability.","excerpt":"Beyond this point, additional unique activations bring little benefit to reconstruction or interpretability","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"The authors recommend choosing the smallest buffer that reaches the quality plateau and replaying it for the remaining training budget.","excerpt":"choose the smallest buffer that reaches the quality plateau and replay it for the remaining budget.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"In their experiments, the authors report reducing activation storage by 8–64× with negligible quality loss.","excerpt":"In our experiments, this reduces activation storage by $8\\text{--}64\\times$ with negligible quality loss.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"},{"text":"The authors report testing the tradeoff across dictionary sizes, token budgets, model layers and sizes, SAE types, and activation sources.","excerpt":"We validate it across dictionary sizes and token budgets, early, middle, and late layers of Llama-3.1, Qwen3 model sizes, and BatchTopK, TopK, and JumpReLU SAEs.","url":"https://arxiv.org/abs/2611.07756","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.07756","checkedAt":"2026-11-08T18:50:08.498Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does the quality plateau change when the activation source changes?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"How Much Data SAE Training Needs","summary":"Authors Nikita Koriagin, Nikita Balagansky and Daniil Gavrilov study how much activation data sparse autoencoder (SAE) training needs when its total training budget is fixed. They report that quality improves as the activation buffer grows, then reaches a plateau: replaying that buffer is nearly as effective as using fresh activations throughout training. Beyond the plateau, they say, adding unique activations brings little benefit to reconstruction or interpretability. The authors describe this diversity-versus-repetition tradeoff with a scaling law for limited data and report testing it across model layers and sizes, SAE types, token budgets and activation sources. Their practical rule is to use the smallest buffer that reaches the quality plateau, then replay it for the remaining budget. In their experiments, this reduced activation storage by 8–64× with negligible quality loss. The result could help reduce a storage bottleneck in SAE training, but the reported savings and quality outcome are experimental findings, not a guarantee for other settings.","question":"How does the quality plateau change when the activation source changes?","claims":[{"text":"The authors report that SAE quality improves as the activation buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","excerpt":"Quality improves rapidly as the buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training.","support":"source-stated"},{"text":"The authors say additional unique activations beyond the plateau bring little benefit to reconstruction or interpretability.","excerpt":"Beyond this point, additional unique activations bring little benefit to reconstruction or interpretability","support":"source-stated"},{"text":"The authors recommend choosing the smallest buffer that reaches the quality plateau and replaying it for the remaining training budget.","excerpt":"choose the smallest buffer that reaches the quality plateau and replay it for the remaining budget.","support":"source-stated"},{"text":"In their experiments, the authors report reducing activation storage by 8–64× with negligible quality loss.","excerpt":"In our experiments, this reduces activation storage by $8\\text{--}64\\times$ with negligible quality loss.","support":"source-stated"},{"text":"The authors report testing the tradeoff across dictionary sizes, token budgets, model layers and sizes, SAE types, and activation sources.","excerpt":"We validate it across dictionary sizes and token budgets, early, middle, and late layers of Llama-3.1, Qwen3 model sizes, and BatchTopK, TopK, and JumpReLU SAEs.","support":"source-stated"}],"limitations":["The supplied abstract gives no numeric buffer sizes or per-condition results, so it does not show how the reported plateau or storage reduction varies across the tested settings."],"mediaId":"1d462f6593cfc99d347bfb4bce73e1ae67fc01471754d1880422eedc33d5485e","correctionReason":null,"id":"f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7","publicationId":"1e4588f7-3a93-4627-8fec-168b41da258a","version":1,"sourceUrl":"https://arxiv.org/abs/2611.07756","sourceTitle":"Scaling Laws for SAE Training Data","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-05","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:00:59.872Z","createdAt":"2026-11-08T17:46:56.079Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"1d462f6593cfc99d347bfb4bce73e1ae67fc01471754d1880422eedc33d5485e","kind":"article","url":"https://arxiv.org/abs/2611.07756","sourceUrl":"https://arxiv.org/abs/2611.07756","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-08T17:46:15.101Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.07756","checkedAt":"2026-11-08T18:50:08.498Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"1d462f6593cfc99d347bfb4bce73e1ae67fc01471754d1880422eedc33d5485e","confidence":0.8700000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"4dbec366c1a85999662762ed6dfa7f499d2339596c9f68a3dd62eb6bf86ca590","decidedAt":"2026-11-08T17:46:15.672Z"},"provenance":{"textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observationId":"1d57713265820b7b1359184088c4e2708a09ba79c33cd54ea6e83f3edd361f33","retrievedAt":"2026-11-08T17:46:15.004Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"3a9c6344-5ffe-4bd1-871a-e4d888faca3e"},"reviewer":{"model":"gpt-6-luna","runId":"3ac55bec-dc3d-4715-822e-32476d5145d3"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T17:46:56.079Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Koriagin, Nikita","url":null,"kind":"unknown"},{"name":"Balagansky, Nikita","url":null,"kind":"unknown"},{"name":"Gavrilov, Daniil","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Koriagin, Nikita","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Koriagin, Nikita","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.07756","locator":"head/meta[name=author|citation_author]"},"assertionId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T18:50:08.498Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","sourceUrl":"https://arxiv.org/abs/2611.07756","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z"}},{"actor":{"name":"Balagansky, Nikita","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Balagansky, Nikita","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.07756","locator":"head/meta[name=author|citation_author]"},"assertionId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T18:50:08.498Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","sourceUrl":"https://arxiv.org/abs/2611.07756","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z"}},{"actor":{"name":"Gavrilov, Daniil","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gavrilov, Daniil","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.07756","locator":"head/meta[name=author|citation_author]"},"assertionId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T18:50:08.498Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","sourceUrl":"https://arxiv.org/abs/2611.07756","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z"}}],"participationCoverage":{"total":3,"returned":3,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.07756","snapshotId":"2d1174fff6ee818185ab873ccc8274327510e9ef8f2d206c2cb5dc2d3eb3b690","contentSha256":"7770c633978d514474971c9a917642303a3445cf1c7643e6654ecb9edd5a3499","textHash":"7cebd910327edf450b7058546d3f48ef01d6dbdf31e5d205fd533f4752ea9f8d","observedAt":"2026-11-08T18:50:08.498Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/f287900f63f94d5dccfc575a7fff5663ce513ac0cbac69cf847a1785869e5ff7/source"}}}],"audits":[{"publicationId":"1e4588f7-3a93-4627-8fec-168b41da258a","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly attributes the study to Nikita Koriagin, Nikita Balagansky and Daniil Gavrilov. It frames the findings, rule and storage savings as the authors' reports rather than established facts.","excerpt":"Authors:Nikita Koriagin, Nikita Balagansky, Daniil Gavrilov"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful to the abstract, but central jargon is unexplained for non-specialists: 'sparse autoencoder (SAE)', 'activations', 'activation buffer', 'replaying', 'interpretability' and 'scaling law'. A reader is not told that SAEs are tools for inspecting what happens inside language models, so the significance is unclear. Smallest repair: add a plain-language gloss of SAE and activations (or term-library tooltips) and one sentence on why SAE training matters. The title and question are otherwise fine; the question follows from the source's activation-source sweeps, whose results the abstract does not report.","excerpt":"Activation storage is often a bottleneck in sparse autoencoder (SAE) training."},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-11-05 matches the arXiv v1 submission date. The feed date of Nov 6 is a later announcement.","excerpt":"[v1]\nThu, 5 Nov 2026 10:00:55 UTC (89 KB)"},{"dimension":"support","status":"pass","explanation":"Every claim matches the abstract, including activation sources, which are covered by the math text and Python code sweeps. The storage-bottleneck significance is grounded in the abstract. The 8–64x figure is scoped to the authors' experiments and is caveated as not generalizing, so nothing is overstated. The limitation note accurately says the abstract lacks per-condition numbers.","excerpt":"Sweeps using mathematical text and Python code test the effect of changing the activation source."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page for this paper. It is the direct original source and the most relevant link.","excerpt":null}],"createdAt":"2026-11-08T17:46:56.079Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}